Implicit Task Composition (ITC) Overview

• ITC is a framework that defines new tasks by reusing and fusing learned components—such as adapters, prompts, or Q-functions—without the need for full model retraining.
• It employs diverse compositional regimes, including algebraic, geometric, and attention-based methods, to integrate information across reinforcement learning, modular transfer, and generative models.
• Empirical studies show ITC’s potential for zero-shot synthesis and safety-aware behavior, though challenges remain in domain narrowness, scalability, and supervision dependency.

Searching arXiv for papers on Implicit Task Composition and closely related task-composition methods. Implicit Task Composition (ITC) denotes a class of methods and problem settings in which a system produces behavior for a composite task by reusing and combining previously learned task representations, skills, prompts, value functions, or modules, often without retraining the full model for each new composition. Across the literature, the term spans several distinct technical regimes: inference-time composition of generative objectives in diffusion models (Lovelace et al., 17 Sep 2025), deployment-time algebra over learned value functions for Boolean reach/avoid tasks (Leahy et al., 2023), weight-space or module-space composition for transfer learning (Zhang et al., 2024, Pfeiffer et al., 2020, Wang et al., 2023), prompt-space composition in frozen visual-LLMs (Perera et al., 2023), and latent or contextual composition in transformers and algorithmic reasoning systems (Kobayashi et al., 2024, Lee et al., 16 Jun 2025, Sindhi et al., 2024). A closely related line studies hidden or unintended goals emerging from composition itself, including latent user subgoals under model mismatch (Tulli et al., 29 Jan 2025) and composition-induced implicit intents in LLM skill marketplaces (Hu et al., 2 Jul 2026). One purported source, "Generalization of Compositional Tasks with Logical Specification via Implicit Planning" (Xu et al., 2024), does not in fact contain scientific content on ITC, because the supplied document is a publisher formatting template rather than a research article.

1. Conceptual scope and problem formulations

ITC is not a single formalism. In reinforcement learning and control, it commonly refers to producing a new policy or value function for a task that was not trained monolithically, but is instead assembled from previously learned task-conditioned objects. In language and vision, it may refer to composing prompts, adapters, or task vectors learned independently, again to obtain a model for a new task without full retraining. In in-context learning, the term is naturally extended to cases where the model appears to infer and execute a composite task from contextual evidence about subtasks, rather than from explicit symbolic decomposition (Lee et al., 16 Jun 2025).

One broad axis of variation is what is being composed. In value-function composition for safe Boolean RL, the learned object is an extended Q-function $\bar{Q}(s,g,a)$ conditioned on state, goal region, and action, and composition is performed by Boolean operators such as pointwise $\min$, $\max$, and an analytic negation rule (Leahy et al., 2023). In task-vector methods, the compositional object is a fine-tuning displacement $\tau_i = \theta_i - \theta_0$ from a shared pretrained model, with new models produced by combining scaled parameter blocks from multiple task vectors (Zhang et al., 2024). In AdapterFusion and AdapterDistillation, the compositional objects are task-specific adapters inserted into a frozen transformer, and composition is performed either dynamically by attention over adapters or offline by distillation into a single student adapter (Pfeiffer et al., 2020, Wang et al., 2023). In prompt algebra, the compositional object is a soft prompt $v_j$ learned for a task, combined linearly as $f(\mathbf{v}) = \sum_i \theta_i v_i$ (Perera et al., 2023). In SpeechOp, ITC is instantiated directly as inference-time composition of denoising score functions for enhancement and text guidance in a latent diffusion model (Lovelace et al., 17 Sep 2025).

A second axis is whether composition is explicit or implicit. Prompt algebra is explicit at the prompt level, because the combined prompt is constructed directly by linear combination, but the resulting classifier behavior over a union or product label space is implicit in the model’s response to that combined prompt (Perera et al., 2023). AdapterFusion stores task knowledge explicitly in separate adapters, yet learns an input-dependent, layer-wise fusion policy that is latent in activations (Pfeiffer et al., 2020). Transformer constructions for modular computation can be explicit in the task-conditioning signal but implicit in the hidden-state execution pathway, as in the case where a latent vector $z$ is provided and multi-head attention internally assembles the effective operator $\omega(z)=\sum_m z^{(m)}\Theta^{(m)}$ (Kobayashi et al., 2024). SpeechOp’s ITC is implicit in a stronger sense: transcript guidance is not trained as an explicit joint enhancement task but emerges from score composition at sampling time (Lovelace et al., 17 Sep 2025).

A third axis concerns when composition occurs. AdapterDistillation composes knowledge during training and deploys only a single distilled adapter at inference, so the final composition is compiled into parameters rather than executed online (Wang et al., 2023). By contrast, AdapterFusion composes at inference time by attending over frozen task adapters (Pfeiffer et al., 2020). Safety-aware Boolean composition in RL is also deployment-time: once primitive extended Q-functions are learned, conjunction and disjunction are formed by pointwise operations without retraining the composite task (Leahy et al., 2023). SpeechOp’s ITC is again inference-time, because the composition is enacted during reverse diffusion through a task-composed score function (Lovelace et al., 17 Sep 2025).

These differences imply that ITC is best understood as a family resemblance concept centered on reuse, modularity, and post-hoc construction of new task behavior.

2. Representation regimes for composition

A useful organizing principle is the representational substrate on which composition is defined.

Regime	Compositional object	Representative papers
Value-function / policy space	Extended Q-functions, soft Q-functions, policies	(Leahy et al., 2023, Rietz et al., 2022)
Module / adapter space	Task adapters and fusion weights	(Pfeiffer et al., 2020, Wang et al., 2023)
Weight space	Task vectors and blockwise scaling coefficients	(Zhang et al., 2024)
Prompt space	Soft prompts in a frozen VLM	(Perera et al., 2023)
Context / hidden-state space	In-context subtask representations, latent task codes	(Lee et al., 16 Jun 2025, Kobayashi et al., 2024)
Score-function space	Conditional diffusion scores	(Lovelace et al., 17 Sep 2025)

In RL, the value-function regime provides some of the cleanest exact composition rules. In safety-aware Boolean composition, tasks are defined over a labeled MDP $\langle \mathcal{S},\mathcal{A},\tau,R,L\rangle$ with formula syntax $$\phi \coloneqq p \mid \neg \phi \mid \phi_1 \wedge \phi_2 \mid \phi_1 \vee \phi_2$$, and composition is performed over optimal extended Q-functions by $\min$0 and $\min$1 (Leahy et al., 2023). This yields exact zero-shot composition for conjunction and disjunction under the paper’s assumptions.

In modular transformer transfer, AdapterFusion separates task learning from composition. Each task obtains its own adapter $\min$2 in a frozen pretrained model, and a target-specific fusion module $\min$3 later combines frozen adapter outputs $\min$4 through an attention-like mixer computed from per-layer query, key, and value matrices (Pfeiffer et al., 2020). Composition is thus a routing problem over reusable modules rather than a re-estimation of the backbone.

In weight-space ITC, aTLAS treats each task vector as a block-structured object $\min$5 and learns a block-diagonal scaling operator $\min$6. New task models are formed as $\min$7, with only the scalar coefficients $\min$8 optimized (Zhang et al., 2024). This suggests a form of composition over low-dimensional subspaces spanned by prior task updates.

In prompt-space ITC, linear prompt algebra is combined with geometric constraints. Independently learned prompts are projected into a low-dimensional subspace derived from the pretrained vocabulary, and then composed as $\min$9 (Perera et al., 2023). The projection is designed to preserve compatibility between prompts by grounding them in the original text embedding geometry.

In in-context settings, the object of composition is often less explicit. In the modular exponential study, transformers receive blocked contexts containing two single-exponential subtasks and then a double-exponential task $\max$0, and the central question is whether the model uses the earlier blocks to infer and execute the composed mapping rather than learning it monolithically (Lee et al., 16 Jun 2025). In the linear-attention construction, the composition is made explicit as $\max$1, but the transformer implements that composition internally through attention heads and an MLP (Kobayashi et al., 2024).

Finally, in SpeechOp the compositional object is a score field over diffusion latents. Enhancement and TTS are combined not by averaging generated samples but by composing their denoising scores in a classifier-free-guidance-like form (Lovelace et al., 17 Sep 2025).

3. Core mechanisms

Different ITC systems can be compared in terms of the operator that turns component tasks into a composite task.

A simple family uses algebraic pointwise operators. Safety-aware Boolean RL employs $\max$2 and $\max$3 over extended Q-functions, plus analytic negation for minimum-violation semantics (Leahy et al., 2023). This is an exact and interpretable composition rule, but it is tied to labeled reachability tasks and strong assumptions such as deterministic MDPs for the theoretical guarantees.

A second family uses geometry-mediated weighting. In task-prioritized policy composition, the composed soft Q-function is $\max$4, where the weight depends on the Jacobian and Hessian of the high-priority Q-function: $\max$5 (Rietz et al., 2022). This makes lower-priority behavior active only in the approximate indifference-space of the higher-priority task. The mechanism is more implicit than constant reward mixing, but remains hand-designed and derivative-dependent.

A third family uses learned routing. AdapterFusion computes input-dependent fusion weights over task adapters at each layer and token position. The output of the feed-forward layer $\max$6 queries the set of adapter outputs $\max$7, and the softmax-weighted mixture determines the fused representation used by the target task (Pfeiffer et al., 2020). This supports contextual, selective borrowing from source tasks while preserving the underlying adapters.

A fourth family uses composition by distillation. AdapterDistillation first learns a new-task adapter locally, then distills the outputs of existing teacher adapters—fused through AdapterFusion-style attention—into a single student adapter (Wang et al., 2023). Composition is therefore explicit during training but implicit at deployment, since only one adapter remains.

A fifth family uses linear combination with structural constraints. Prompt algebra composes learned prompts linearly and improves compositionality by constraining prompt vectors to a vocabulary-derived eigenspace and regularizing prompted predictions so they preserve semantic relations already present in the frozen VLM (Perera et al., 2023). aTLAS similarly uses linear combination, but in parameter space and with anisotropic blockwise coefficients that can differ across tasks and parameter groups (Zhang et al., 2024).

A sixth family uses hidden-state or attention-based assembly. In the constructive transformer result, two tokens $\max$8 and

$\max$9

are processed by a linear-attention block whose $\tau_i = \theta_i - \theta_0$0 heads compute and weight the module outputs $\tau_i = \theta_i - \theta_0$1 by $\tau_i = \theta_i - \theta_0$2, writing $\tau_i = \theta_i - \theta_0$3 into the residual stream before an MLP applies $\tau_i = \theta_i - \theta_0$4 (Kobayashi et al., 2024). This offers a mechanistic example of latent modular composition in a transformer.

A seventh family uses score composition in generative models. SpeechOp derives the target score for transcript-guided enhancement as $\tau_i = \theta_i - \theta_0$5 and approximates it by

$\tau_i = \theta_i - \theta_0$6

The paper argues this is superior to naive score averaging because it adds a discriminative transcript-consistency term rather than importing the full unconditional TTS prior (Lovelace et al., 17 Sep 2025).

4. Empirical patterns across domains

The strongest evidence for ITC varies by domain and representation type.

In safe Boolean RL, the main empirical contribution is qualitative and task-specific. The framework demonstrates zero-shot conjunction and disjunction over learned tasks in grid worlds, image-based DQN environments, and a continuous-action Bullet-Safety-Gym setup, while extending Boolean composition to safety semantics such as minimum-violation and prioritized safety (Leahy et al., 2023). The paper states that minimum-violation and prioritized safety produce different path preferences, and that continuous-action composition is achieved by selecting between constituent actor outputs rather than by exact maximization over the composed critic.

In adapter-based NLP transfer, AdapterFusion reports average accuracy 77.33 over 16 NLU tasks, compared with 76.05 for single-task adapters and 75.51 for full fine-tuning, with especially large gains on low-resource tasks such as RTE and MRPC (Pfeiffer et al., 2020). The analysis shows that some tasks rely heavily on their own adapters, whereas small-data tasks borrow strongly from large-source-task adapters such as MNLI and QQP, indicating that learned module composition is selective rather than uniform.

AdapterDistillation is motivated by the scalability limitations of explicit runtime fusion. In FAQ retrieval for task-oriented dialogue, it achieves average accuracy 86.04 versus 85.85 for AdapterFusion, while retaining only adapter-level parameter overhead at inference. Under a 100 GB storage budget, the reported tenant capacities are 261 for full fine-tuning, 1161 for AdapterFusion, and 17587 for AdapterDistillation (Wang et al., 2023). This suggests that offline composition-by-distillation can preserve much of the transfer benefit while eliminating the runtime cost of explicit module fusion.

Prompt algebra shows that independently learned prompts can support post-hoc composition in a frozen CLIP-like model. On UT-Zappos, class-agnostic regularized prompt algebra reports attribute accuracy 46.58, object accuracy 48.68, seen pair accuracy 25.61, unseen pair accuracy 53.46, and AUC 9.48, with the paper stating that the composite model improves classification accuracy over the best base model by 8.45% on average on this dataset (Perera et al., 2023). On CIFAR10, prompt algebra improves overall union-task accuracy to 96.29 or 96.64 depending on regularization, exceeding the best base prompt in the reported setting (Perera et al., 2023). These results support the narrower claim that prompt-space composition can recover a balanced classifier over union and product label spaces without retraining the backbone.

The in-context modular exponential study supplies some of the clearest behavioral evidence for latent composition in transformers. Models trained with blocked subtask curricula show zero-shot inference on the composed task, stronger decoding of intermediate variables such as $\tau_i = \theta_i - \theta_0$7 and $\tau_i = \theta_i - \theta_0$8 during the compositional block, and failure under mismatch experiments where the subtask blocks and compositional block use inconsistent parameters (Lee et al., 16 Jun 2025). This combination of zero-shot behavior, probing evidence, and mismatch sensitivity suggests that at least some transformers learn to use contextual subtask structure compositionally rather than merely fitting the final mapping.

SpeechOp provides a direct, named ITC result. In speech enhancement, plain SpeechOp without transcript guidance attains WER 8.1%, whereas SpeechOp + ITC reduces this to 2.9%; the paper highlights this as a 46% relative reduction versus HiFi-GAN-2, which reports WER 5.4 (Lovelace et al., 17 Sep 2025). The TC-CFG ablation with gold transcripts further shows that task-composed classifier-free guidance outperforms naive score averaging, reducing WER from 3.4 to 2.1 while improving PESQ and SpeechBERTScore (Lovelace et al., 17 Sep 2025). This is one of the strongest available demonstrations that inference-time composition can repair a concrete failure mode of single-task generative systems.

5. ITC in relation to hidden goals, safety, and intent emergence

A distinctive branch of the literature concerns task composition that is not requested explicitly, or whose consequences emerge only when multiple components interact.

In alignment under model mismatch, implicit goals are modeled as bottleneck states in the user’s task model that the user omits because they appear unavoidable. The robot’s task is therefore effectively a composition of the explicit goal set $\tau_i = \theta_i - \theta_0$9 with a latent set of intermediate state-visit requirements $v_j$0 (Tulli et al., 29 Jan 2025). The method identifies bottleneck candidates across a set of possible human models, computes maximal achievable subsets in the robot’s model, and constructs an optimal query MDP to minimize the number of clarification queries needed to identify a guaranteed policy. This is narrower than general ITC, but it provides a formal account of how an explicit task can need to be augmented by implicit conjunctive subgoals.

In LLM skill marketplaces, composition itself can create unintended objectives. SkillFuzz defines an implicit intent as an unintended objective $v_j$1 that appears in the agent’s plan under a skill activation vector $v_j$2, but is not required by the user task $v_j$3 nor prescribed by any individual skill (Hu et al., 2 Jul 2026). The paper formulates implicit-intent discovery as fuzzing over skill compositions, using plan drift from the skill-free baseline as a differential oracle. Across representative workloads, it reports over 1,000 distinct implicit intents under fixed query budgets, with more than 80% of the highest-risk flagged compositions confirmed during execution-time validation (Hu et al., 2 Jul 2026). The central ITC implication is that composition is not merely a mechanism for benign reuse; it can induce new effective tasks, including high-severity ones, that are absent from every constituent skill in isolation.

These papers complicate the standard view of ITC as constructive reuse. They suggest a broader interpretation in which composition is also a source of latent constraints, emergent objectives, and safety failures.

6. Open issues, limitations, and boundaries of the concept

Despite the diversity of methods, several recurring limitations define the current state of ITC.

One limitation is domain narrowness. Boolean value-function composition is exact but limited to labeled reach/avoid tasks and depends on assumptions such as deterministic MDPs and shared dynamics (Leahy et al., 2023). Prompt algebra is compelling but restricted to prompt-tuned frozen VLMs and mostly two-task compositions over classification tasks (Perera et al., 2023). SpeechOp’s ITC is highly effective for transcript-guided enhancement, but it is a specific score-composition strategy inside one multi-task latent diffusion architecture (Lovelace et al., 17 Sep 2025).

A second limitation is the gap between expressivity and learnability. The linear-attention transformer note shows that transformers can in principle implement modular latent composition of the form $v_j$4, yet it provides no empirical evidence that gradient-trained models actually discover such decompositions from data (Kobayashi et al., 2024). The contextual curriculum study goes further empirically, but still in a synthetic algorithmic domain (Lee et al., 16 Jun 2025).

A third limitation is compositional closure. Prioritized safety in RL is not closed under unrestricted implicit composition because each learned value function may only store its own optimal path structure; the paper identifies chattering and proposes richer extended value functions as a remedy, at increased complexity (Leahy et al., 2023). In prioritized policy composition via indifference-space, the method offers heuristic structure and strong intuitions about constraint preservation but no formal theorem guaranteeing exact high-priority satisfaction or convergence (Rietz et al., 2022).

A fourth limitation is supervision dependence. AdapterFusion composes modules implicitly, but only after target-task supervised training of the fusion weights (Pfeiffer et al., 2020). aTLAS often learns task-vector coefficients from at least a small amount of target data, even though the source task-vector bank is reusable (Zhang et al., 2024). AdapterDistillation likewise requires local labeled data for the new task (Wang et al., 2023). These methods are ITC-adjacent, but not zero-shot universal composition operators.

A fifth limitation is scalability of structured composition search. The task-agnostic architecture proposal frames ITC as internal execution and composition of prior skills via a stateful Transformer-like virtual machine, with recursive composition equations such as

$v_j$5

yet it remains a position paper with no implementation or empirical validation (Sindhi et al., 2024). SkillFuzz addresses combinatorial explosion with contract-guided Monte Carlo tree search, but its search depth is capped and its oracle depends on the availability of plan artifacts (Hu et al., 2 Jul 2026).

A sixth and more basic boundary concerns evidence quality. The supplied source for (Xu et al., 2024) does not support any claims about compositional RL, logical specifications, latent planning, or ITC, because it is an IEEE Transactions on Artificial Intelligence formatting template rather than a scientific paper. Any substantive account of that document as an ITC contribution would therefore be unsupported.

These limitations suggest that the field still lacks a unified theory of task composition across representation spaces, as well as broad empirical evidence that the same mechanisms scale from synthetic or structured settings to open-ended ones.

7. Significance and emerging synthesis

Taken together, the literature suggests that ITC is best defined operationally rather than taxonomically. A method belongs to the ITC family when it uses reusable task-bearing objects to produce behavior for a new task through composition rather than full end-to-end retraining. Those task-bearing objects may be value functions, policies, prompts, adapters, task vectors, latent task codes, score functions, or contextual subtask traces. The composition operator may be algebraic, geometric, attention-based, distillative, or guidance-based. The composite task may be benignly intended, as in transcript-guided speech enhancement (Lovelace et al., 17 Sep 2025) or safe Boolean goal composition (Leahy et al., 2023), or it may emerge implicitly and undesirably, as in marketplace-scale skill interactions (Hu et al., 2 Jul 2026).

A plausible synthesis is that ITC research is converging on three general design principles. First, reusable task structure must be represented in a form that remains separable enough to compose later, whether by freezing adapters (Pfeiffer et al., 2020), storing task vectors relative to a common base model (Zhang et al., 2024), or maintaining primitive extended Q-functions over goals (Leahy et al., 2023). Second, the composition operator must control interference, either by exact semantics such as min/max, by anisotropic scaling, by input-dependent routing, or by discriminative guidance rather than naive prior mixing (Leahy et al., 2023, Zhang et al., 2024, Lovelace et al., 17 Sep 2025). Third, useful composition often depends on preserving an intermediate factorization between task inference and task execution, as suggested by curriculum-induced in-context strategies (Lee et al., 16 Jun 2025) and by constructive latent modularity in transformers (Kobayashi et al., 2024).

This suggests that “Implicit Task Composition” is less a single technique than a research program concerned with how learned task structure can be stored, exposed, combined, and controlled. The central unresolved question is not whether composition is possible in isolated settings, but which representational and optimization choices make it reliable, scalable, and safe across domains.